Estimation of time‐varying autoregressive model parameters by multilayered neural networks: Identification of local stationary sections from noise variance

Abstract

This paper proposes a method of estimating time-varying autoregressive (AR) model parameters by means of multilayered neural networks (NNs). First, an estimation method for time-varying AR parameters is discussed under the assumption of local stationarity. Then it is shown that the section length T satisfying local stationarity is very important in maintaining the balance of estimation precision for both noise variance and time-varying AR parameters. Next, on the basis of the nonlinear Yule-Walker equation, the noise variance is estimated by NNs. The section length T satisfying local stationarity is determined by estimated noise variance, and for the specified section length T, time-varying AR parameters are estimated by the Yule-Walker method. Finally, from numerical simulation results it is shown that the proposed method is very effective for the estimation of time-varying AR model parameters. © 1997 Scripta Technica, Inc. Electron Comm Jpn Pt 3, 80(7): 20–27, 1997